What an eighteenth-century hand makes an OCR system prove

A scanner can preserve an old page without making its writing easy to read. The National Archives uses a 1789 letter to absent senators to teach the difference between seeing marks and interpreting them.

The letter is headed New York, March 11th, 1789. In its first body paragraph appears the phrase “the house of Representatives.” The long s in the historical hand resembles an f. The Archives’ teaching example renders the tempting reading as “houfe of Reprefentatives,” then explains that those shapes represent s in the original spelling. The point is not that the page is defective. The point is that the reader needs knowledge of the writer’s letterforms and period conventions.

The Archives’ transcription guidance gives a useful discipline for an AI-assisted workflow. Retain original spelling, abbreviations, and punctuation. Compare uncertain letters with other letters by the same writer. If a word cannot be resolved, mark it as [illegible] rather than silently repairing it. Tags and comments can make a transcription searchable without silently altering the historical wording.

Imagine a hypothetical scan review. An OCR system produces “houfe of Reprefentatives.” That output is not evidence that the document used f in those words. A reviewer compares the long shape with other instances of s in the same hand and records the reading as “house of Representatives” while noting that the source uses the historical long-s glyph. Reading the glyph as s does not modernize the spelling. It identifies the letter represented by the historical form. If the comparison remains uncertain, the uncertainty belongs in the record.

A separate modernized aid could regularize spelling for search or general readers, but it should be labeled as a modernized layer rather than substituted for the transcription. The original spelling and punctuation remain visible. That distinction matters because a smooth correction can erase the very feature a historian may want to study.

The practical lesson is modest. Ask an OCR or AI tool for a candidate transcription, then compare the candidate with the image and with the writer’s repeated letterforms. Keep the original visible. Mark unresolved words. Treat the machine’s confidence as a prompt for review, not as permission to rewrite the source.

The 1789 letter does not prove that every historical hand defeats OCR, and it does not provide a benchmark for modern systems. It gives one concrete reason to slow down: a familiar letter can carry an unfamiliar historical shape.